Serial Order in Planning the Production of Successive ...ardiroel/JML1996.pdf · JOURNAL OF MEMORY...
Transcript of Serial Order in Planning the Production of Successive ...ardiroel/JML1996.pdf · JOURNAL OF MEMORY...
JOURNAL OF MEMORY AND LANGUAGE 35, 854–876 (1996)ARTICLE NO. 0044
Serial Order in Planning the Production of SuccessiveMorphemes of a Word
ARDI ROELOFS
Max Planck Institute for Psycholinguistics, Nijmegen, The Netherlands
Five implicit priming experiments examined whether the speech production system can plannoninitial morphemes of a word in advance of initial ones. On each trial, subjects had to produceone word out of a set of three words as quickly as possible. In a homogeneous condition, theresponses shared part of their form, whereas in a heterogeneous condition they did not. The firstexperiment shows that the task is sensitive to morphological planning. In producing disyllabicsimple and compound nouns, a larger facilitatory effect was obtained when a shared initialsyllable constituted a morpheme than when it did not. The next three experiments suggest thatsuccessive morphemes are planned in serial order. In producing nominal compounds, no facilita-tion was obtained for noninitial morphemes. In producing prefixed verbs, facilitation was obtainedfor the prefix but not for the noninitial base. Sharing morphemes often implies semantic overlap.The fifth experiment shows that semantic similarity per se yields inhibition rather than facilitation.Computer simulations show that the WEAVER model of word-form encoding (Roelofs, 1992b,1994, submitted-a) accounts for the findings. q 1996 Academic Press, Inc.
Speech production requires advance plan- lyzed in their constituent morphemes. If mor-phemes are not stored with words in memory,ning. This raises the question of what the plan-
ning units are and which constraints hold for they cannot be used in production. Thus, onlydecomposed form entries allow morphemes tothe planning process (e.g., Lashley, 1951; Jor-
dan & Rosenbaum, 1989). For example, what be planning units. In a decomposition view,the morphemes of a word may be plannedare the degrees of freedom in planning poly-
morphemic words? Issues of advance plan- serially or in parallel.This paper addresses the question ofning and lexical representation are related.
Two major views in the psycholinguistic re- whether the speech production system canplan noninitial morphemes of a word beforesearch on morphological complexity are the
full listing hypothesis and the decomposition initial ones. Evidence from speech errors sug-gests that prefixed words such as disappearhypothesis (Butterworth, 1983). According to
the full listing hypothesis, the lexical represen- and reappear are assembled out of their mor-phemes during production (e.g., Dell, 1986;tations of polymorphemic forms are morpho-
logically unanalyzed, whereas according to Stemberger, 1985a). No evidence exists, how-ever, that bears on the question of whether itthe decomposition hypothesis, they are ana-is possible to plan the morpheme appear inadvance if it is known beforehand that either
I am grateful to Maarten van Casteren for his help in the word disappear or the word reappear haspreparing the experiments, running them, and writing the
to be produced. The current paper providesIcon and AWK computer programs for the exhaustiveevidence that successive morphemes of asearch of the CELEX lexical database for stimulus materi-
als; and to Rob Schreuder, Antje Meyer, Pim Levelt, word have to be planned one after another.Nanda Poulisse, Harald Baayen, D. Balota, and three Preparation of noninitial morphemes of aanonymous reviewers for helpful comments and sugges-
word before initial ones is not possible, eventions. Address reprint requests to Ardi Roelofs, Maxwhen these noninitial morphemes correspondPlanck Institute for Psycholinguistics, P.O. Box 310, 6500
AH Nijmegen, The Netherlands. to the forms of fully fledged words.
8540749-596X/96 $18.00Copyright q 1996 by Academic Press, Inc.All rights of reproduction in any form reserved.
AID JML 2453 / a004$$$$81 11-07-96 07:33:33 jmlas AP: JML
855PRODUCTION OF POLYMORPHEMIC WORDS
The work reported in this paper deals with words, both monosyllabic and polysyllabic(i.e., Meyer, 1990, 1991). However, whetherlexical access in speech production, in particu-
lar, the second stage of the access process seriality holds for the encoding of morphemesof polymorphemic words is unknown. In gen-involving the encoding of word forms (the
first stage is lemma retrieval and will not be eral, whereas morphological complexity hasreceived much attention in the study of lan-addressed). Theories conceive of word-form
encoding (e.g., Dell, 1986, 1988; Levelt, 1989, guage comprehension (e.g., Henderson, 1985;Marslen-Wilson, Tyler, Waksler, & Older,1992) as the mapping of a representation of
the word as a semantic–syntactic entity (the 1994; Schriefers, Zwitserlood, & Roelofs,1991) and language acquisition (especially in-word’s lemma) onto an articulatory program
(phonetic plan). This is achieved by recov- flectional morphology, e.g., Pinker & Prince,1988), it has been largely ignored in the studyering the word’s morphemes and phonemic
segments from memory and serially grouping of language production. The comprehensionbias in psycholinguistics may be due to a lackthe segments into phonological syllables.
These syllables are used to derive the articula- of appropriate experimental techniques forproduction research (cf. Meyer, 1992). I willtory program. The phonological syllables cre-
ated during the encoding process together argue that it is important to study morphologyin speech production. Furthermore, it will be-make up phonological words (e.g., Levelt,
1989, 1992; Levelt & Wheeldon, 1994). Pho- come clear from the research reported herethat available techniques are sensitive to mor-nological words correspond to domains of syl-
labification and of stress assignment and are phological structure in production.It is important to examine whether serialitydomains for the application of phonological
rules (e.g., Booij, 1983, 1995; McCarthy & holds for polymorphemic words, because itseems plausible that left-to-right encodingPrince, 1990, 1993). Phonological words may
be smaller than lexical words (e.g., appear in does not hold for morphemes. Morphemes of-ten constitute fully fledged phonologicaldisappear) or larger (e.g., the cliticized form
appearin combining the verb appear and the words of their own. This means that informa-tion about the form of these morphemes is, topreposition in).
Theories of the encoding of word forms often a large extent, independent of the informationabout other morphemes in the word (e.g.,assume that a word is planned in a rightward
incremental fashion (e.g., Dell, 1986, 1988; Goldsmith, 1990; Spencer, 1991). Thus, itseems plausible that they function as indepen-Kempen & Hoenkamp, 1987; Levelt, 1989,
1992). Incremental planning means that an en- dent planning units as far as serial order isconcerned.coding stage is initiated by a critical fragment
of the output of a preceding stage rather than The paper is organized as follows. First, toset the theoretical scene, I briefly describe theits complete output. That is, a process starts
working on the basis of partial input. For exam- WEAVER model of word-form encoding inspeech production (Roelofs, 1992b, 1994, sub-ple, syllabification starts when it receives the
initial segments of a word. The process does not mitted-a). WEAVER (Word-form Encoding byActivation and VERification) is a computerhave to wait until all the segments of the word
have been made available. A central assumption model developed within the theoretical frame-work for speech production advanced by Leveltof these theories is that processing proceeds in
a ‘‘rightward’’ fashion, that is, from the begin- and colleagues (e.g., Levelt, 1989, 1992; Le-velt & Wheeldon, 1994). The model sharesning of a word to its end (i.e., Dell, 1988; Levelt,
1989, 1992). much in common with other approaches (e.g.,Dell, 1986, 1988). WEAVER is taken as theThe assumption of rightward incremental-
ity is supported by a variety of empirical find- theoretical framework for the current paper be-cause it is more explicit than other approachesings about the encoding of monomorphemic
AID JML 2453 / a004$$$$82 11-07-96 07:33:33 jmlas AP: JML
856 ARDI ROELOFS
about the particular issues addressed. Second, I stead of the imperatival form [bə.hal], themood parameter has to be set. The lemma anddescribe the experimental paradigm used: the
implicit priming paradigm developed by Meyer its diacritic features are input to word-formencoding. The articulatory program is derived(1990, 1991). Third, I report the results of five
experiments testing the model. Finally, I will in three major steps: morphological encoding,phonological encoding, and phonetic encod-show by computer simulation that WEAVER
accounts for the findings. The paper ends with ing (Levelt, 1989). The ‘‘morphological en-coder’’ takes the lemma of behalen plus itsa general discussion addressing a number of em-
pirical and theoretical issues raised by the exper- diacritics, and outputs, respectively, the prefix,root, and plural suffix morphemesiments.
THEORETICAL FRAMEWORK»be…, »haal…, and »en….
Following Levelt and colleagues, Weaverconceives of lexical access in speech produc- This process thus concerns what is tradition-tion as a process consisting of two major steps, ally called the ‘‘syntax-morphology inter-called lemma retrieval and word-form encod- face’’ (e.g., Spencer, 1991). The ‘‘phonologi-ing (cf. Dell, 1986). In conceptually driven cal encoder’’ successively takes »be…, »haal…,lemma retrieval, a lexical concept is used to and »en… and produces the phonological wordrecover the lemma of the corresponding word (skipping the foot level)from memory. A lemma is a memory repre-sentation of the syntactic properties of a word.For example, a verb lemma says that the wordis a verb and makes explicit the word’s argu-ment structure. A verb lemma also contains anumber of abstract morphosyntactic slots forthe specification of tense and agreement pa-
PhWd
PhWd
b
s
e eh a l
ss sw
rameters and for the specification of mood(e.g., indicative, imperative). Setting these pa- That is, the process delivers a syllabified se-
quence of segments, together with a stress pat-rameters provides an index to a form pointer.The lemma retrieval process delivers this tern over the syllables (s), where s and w
stands for metrically strong and weak (cf. Li-pointer to the processes that recover theword’s morphophonological properties from berman & Prince, 1977). This representation
says that the infinitival form of behalen consti-the mental lexicon and that encode the formof the word. The memory representation of tutes a phonological word (PhWd) comprising
a syllable corresponding to the prefix and anthese form properties is sometimes called theword’s lexeme (e.g., Kempen & Huijbers, embedded phonological word corresponding
to the base (cf. Booij, 1983, 1995; McCar-1983).Assume a Dutch speaker wants to verbalize thy & Prince, 1990, 1993).1 The prefix syllable
has /b/ as onset and /ə/ as nucleus. The embed-the concept TO OBTAIN. First, the ‘‘lemmaretriever’’ takes the lexical concept and makes ded phonological word consists of two sylla-
bles. The first syllable has /h/ as onset andavailable the lemma information of the wordbehalen (for a theory and computer model of /a/ as nucleus, and the second syllable has /l/
as onset and /ə/ as nucleus. The process thatlemma retrieval, see Roelofs, 1992a, 1992b,1993). That is, the process delivers the syntac-tic property verb, slots for the word’s mood,
1 I assume that recursion of PhWd is possible. Foottense, and agreement values, and a form and syllable theories, however, exclude recursion of thepointer. To encode the appropriate word form, categories Ft and s (for arguments and references, see
Booij, 1983, 1995; McCarthy & Prince, 1993).for example, the infinitival form [bə.ha.lə] in-
AID JML 2453 / a004$$$$82 11-07-96 07:33:33 jmlas AP: JML
857PRODUCTION OF POLYMORPHEMIC WORDS
generates this phonological word representa- scribes behalen in terms of the syllable pro-grams [bə], [ha], and [lə], which are recoveredtion thus comprises what is traditionally called
the ‘‘morphology-phonology interface’’ (e.g., from a phonetic syllabary. The articulatoryprogram makes explicit the gestural scores forGoldsmith, 1990).
Dutch prefixes such as »be…, »ver…, and »ont… the articulatory movements and indicates,among other things, that the second syllableare independent syllabification domains, but
they are not phonological words of their own should be pronounced louder or longer thanthe other syllables. This encoding stage thus(Booij, 1995). For example, the segment /r/
of »ver… in the verb verachten (despise) is not includes what is sometimes called the ‘‘post-lexical phonology’’ (e.g., Goldsmith, 1990).syllabified with the base verb achten, as the
Maximal Onset Principle (e.g., Goldsmith, The WEAVER model (Roelofs, 1994, sub-mitted-a) computationally implements the en-1990) would predict, but is made the coda of
»ver…. This does not hold for the string vera coding processes just described. WEAVERintegrates a spreading-activation based net-in a simple word such as veranda (verandah),
which is syllabified as (və)s(rɑn)s(dɑ)s . Thus, work with a parallel object-oriented produc-tion system. The type of system is a mix ofin models that assume that syllabifications are
computed rather than stored (e.g., Levelt, traditional AI, connectionism, and traditionalcognitive modeling (cf. Anderson, 1983). The1992; Roelofs, 1992b, 1994, submitted-a), the
syllabification process has to ‘‘know’’ the model conceives of the word-form lexicon asa network of morphophonological nodes andmorphemic source of the segments that it re-
ceives. The process cannot blindly accept a labeled links. The network is accessed byspreading of activation. Activation of nodesstring of segments and syllabify the segments
without taking morpheme boundaries into ac- triggers procedures that build incrementallya phonetic plan. An important task of thesecount. This implies that the lexical entries of
words have to indicate morpheme boundaries. procedures is to verify the link between anactivated node and the selected nodes oneThe prefixes mentioned above are, however,
not phonological words of their own. A pho- level up in the network. Morphological proce-dures select the morpheme nodes that appro-nological word must contain at least one
stressable syllable (phonological words corre- priately encode a selected lemma and its tense,agreement, and mood parameters. Phonologi-spond to domains of stress assignment), but
»ver… and »be… do not. Because prefix syllables cal procedures select the phonemic segmentsof the morphemes and syllabify the segmentsare metrically dependent on a host, they have
to be adjoined to the phonological word corre- in order to construct phonological syllables asconstituents of phonological word representa-sponding to the base verb. For example, the
syllable (bə)s realizing the prefix »be… of beha- tions. Finally, phonetic procedures select thearticulatory programs that appropriately en-len is adjoined to the disyllabic phonological
word ((ha)s(lə)s)PhWd , creating the trisyllabic code these phonological syllables.Elsewhere (Roelofs, 1994, submitted-a), Iphonological word ((bə)s((ha)s(lə)s)PhWd)PhWd .
Finally, the ‘‘phonetic encoder’’ takes the have shown by computer simulation that theWEAVER model accounts for key empiricalphonological word representation delivered
by the phonological encoder and produces the findings about the time course of phonologicalfacilitation and inhibition from spoken dis-articulatory program,tractors in picture naming (Meyer & Schrie-fers, 1991), for effects from the order of en-
[bə][*ha][lə].coding inside and between the syllables of aword (Meyer, 1990, 1991), for effects fromword and syllable frequency (Jescheniak &According to Levelt (1989, 1992; Levelt &
Wheeldon, 1994), this representation de- Levelt, 1994; Levelt & Wheeldon, 1994), and
AID JML 2453 / a004$$$$82 11-07-96 07:33:33 jmlas AP: JML
858 ARDI ROELOFS
for classical speech errors (Nooteboom, come from distinct phrases. This is also char-acteristic of whole-word exchanges (e.g., as1969). Furthermore, novel predictions con-
cerning word and sentence production have in ‘‘we completely forgot to add the LISTto the ROOF,’’ from Garrett, 1980), whichbeen tested and validated in new experiments
(Roelofs, submitted-a, submitted-b). virtually always involve items of the samesyntactic category and typically ignore phraseAn important aspect of the model for the
present paper is that the encoding algorithm boundaries (Garrett, 1975). This suggests thatthese morpheme errors and whole-word errorsprovides for a suspension-resumption mecha-
nism that supports (rightward) incremental occur at the same level of processing. Theyoccur when lemmas in a developing syntacticgeneration of phonetic plans. The three pro-
cessing stages (i.e., morphological encoding, structure trade places. By contrast, the ex-changing morphemes in an error such asphonological encoding, and phonetic encod-
ing) compute aspects of a word form in paral- ‘‘SLICEly THINNed’’ (from Stemberger,1985a) belong to different syntactic categorieslel from the beginning of the word to its end.
When a stage has used the available informa- (adjective and verb) and come from the samephrase. This is also characteristic of segmenttion before reaching the end of the word, it
stops and waits until it gets new input. When exchanges (e.g., as in ‘‘Rack Pat’’ for ‘‘packrat,’’ from Garrett, 1988), which are typicallyfurther information is provided, the stage con-
tinues from where it stopped. not affected by lemma information such assyntactic class and occur on words within a
EVIDENCE FOR MORPHOLOGICAL single phrase. This suggests that this secondSTRUCTURE IN LEXICAL ENTRIES type of morpheme error and segment errors
occur at the same level of processing, namelyThe literature about the morphological pro-cesses and structures underlying speech pro- the level at which lexemes are retrieved and
the morphophonological form of the utteranceduction is scarce. The existing empirical evi-dence mainly comes from two sources: speech is constructed. The errors occur when mor-
phemes or segments in a developing morpho-errors in normal and aphasic speakers and pro-duction latencies obtained in word pronounci- phonological structure trade places.
In the classification of speech errors, a dis-ation tasks. The present discussion will be re-stricted to evidence concerning the types of tinction is made between contextual and non-
contextual errors. Contextual errors involve apolymorphemic words playing a role in thepresent paper, in particular, prefixed words misordering within the intended utterance,
whereas for noncontextual errors there doesand nominal compounds. For reviews of theevidence concerning inflected forms, see not exist a clear source within the utterance.
Morphemes of both prefixed words and com-Stemberger and MacWhinney (1986) andLevelt (1989). pounds are involved in speech errors (all ex-
amples of errors below are from Stemberger,Speech Errors 1985a). Examples of contextual errors involv-
ing prefixes are the anticipation error ‘‘weThe evidence from speech errors concernsfailures in the selection and serial ordering have twenty-five DEdollars deductible . . .’’
for ‘‘we have twenty-five dollars deductibleof morphemes in an utterance. The evidencesuggests that some morphemic errors concern . . . ,’’ the perseveration error ‘‘it does not
explain how an apparent case of rule EXser-the lemma level, whereas others involve thelexeme level (e.g., Dell, 1986; Garrett, 1975, tion may arise’’ for ‘‘it does not explain how
an apparent case of rule insertion may arise,’’1980, 1988). For example, in ‘‘how manyPIEs does it take to make an APPLE?’’ (from and the exchange error ‘‘a self-INstruct DE
. . .’’ for ‘‘a self-destruct instruction.’’ TheseGarrett, 1988), the interacting stems belong tothe same syntactic category (i.e., noun) and errors involve words of different syntactic
AID JML 2453 / a004$$$$82 11-07-96 07:33:33 jmlas AP: JML
859PRODUCTION OF POLYMORPHEMIC WORDS
classes, which suggests that the errors are due morphological derivations in producing thesederived forms. However, a major problemto encoding failures at the lexeme level. Ex-
amples of noncontextual errors involving pre- with this task is that it forces speakers to pro-duce a word from a perceived morphologi-fixes are the substitution error ‘‘she’s so
EXquisitive’’ for ‘‘she’s so inquisitive,’’ cally related word. This may evoke processesand representations that are not at play duringthe addition error ‘‘positively or negatively
REmarked as . . .’’ for ‘‘positively or nega- the normal production of polymorphemicwords in speech, where the production istively marked as . . . ,’’ and the deletion er-
ror ‘‘they weren’t _jeal_’’ for ‘‘they weren’t based on semantic and syntactic features (i.e.,lemma information) rather than on the percep-conjealing.’’ Similar errors involving prefixes
have also been observed for Dutch (e.g., Pou- tion of a base form. Moreover, by presentinga base form it is difficult to obtain an answerlisse, 1989). These errors are difficult to ex-
plain purely in phonological terms, because to the question at stake in this paper, namelywhether the base is encoded before or afterphonological errors rarely involve more than
a single segment or syllable constituent (e.g., the derivational affix.Dell, 1986; Stemberger, 1985a).
THE IMPLICIT PRIMING PARADIGMSpeech error evidence also suggests thatcompounds have internal morphological To avoid some of the limitations associated
with speech-error corpora and a pronouncia-structure in the mental lexicon. Examples ofmisorderings are ‘‘oh, you were just closing tion task, the present series of experiments
employed the ‘‘implicit priming’’ paradigmthe LIDBOXES’’ for ‘‘oh, you were just clos-ing the boxlids’’ and ‘‘did we miss the TURN developed by Meyer (1990, 1991). This para-
digm involves producing words from learnedTRAIL-off?’’ for ‘‘did we miss the trail turn-off?’’ Again, due to the large number of seg- paired-associates. In her experiments, subjects
first learned small sets of word pairs such asments involved, these errors cannot be ex-plained phonologically. lucht–raket, berg–ravijn, and so forth (sky–
rocket, mountain–ravine, etc.); lucht–raket,Although speech errors may bear on therepresentation of morphology in speech pro- klerk–loket, and so forth (sky–rocket, clerk–
ticket-window, etc.); or lucht–raket, rechter–duction, they do not reveal much about thetime course of planning the production of po- bewijs, and so forth (sky–rocket, judge–
proof, etc.). After learning a set, they had tolymorphemic words (cf. Meyer, 1992). Forexample, the error ‘‘SLICEly THINNed’’ produce the second word of a pair (e.g., raket)
upon the visual presentation of the first worddoes not reveal whether the base thin is en-coded before, simultaneously with, or after the (i.e., lucht). The instruction was to respond as
quickly as possible without making mistakes.derivational affix ly. The question of whethermorphemes are planned in serial order calls The production latency (i.e., the interval be-
tween prompt onset and speech onset) wasfor a chronometric technique. One such tech-nique is word pronounciation. the main dependent variable. An experiment
comprised homogeneous and heterogeneousWord Pronounciation Latencies response sets. In a homogeneous set, the re-
sponse words shared part of their form and inIn word pronounciation experiments, sub-jects are presented with words or pseudowords a heterogeneous set they did not. In the exam-
ple, the responses share the first syllable (RA-to pronounce either as they stand or in a mor-phologically derived form. For example, ket, RAvijn, etc.) or the second syllable (ra-
KET, loKET, etc.) or they are unrelated (raket,MacKay (1978) had subjects produce nomi-nalisations from spoken verbs, such as deci- bewijs, etc.). Heterogeneous sets in the experi-
ments were created by regrouping the pairssion from decide. Differences in pronouncia-tion latencies have been taken as evidence for from the homogeneous sets. Therefore, each
AID JML 2453 / a004$$$$83 11-07-96 07:33:33 jmlas AP: JML
860 ARDI ROELOFS
word pair was tested both under the homoge- phemic words. Assume the response set con-sists of the Dutch prefixed verbs behalen (ob-neous and the heterogeneous condition, and
all uncontrolled item effects were kept con- tain), belopen (walk), and beschieten (shootat) sharing the prefix be. Before the beginningstant across these conditions. Meyer found a
facilitatory effect from homogeneity only of a trial, the morphological encoder can planthe first morpheme »be…, but not the subse-when the overlap was from the beginning of
the response words onward. Thus, a facilita- quent morpheme because it is not shared. Thephonological encoder can prepare the firsttory effect was obtained for the set that in-
cluded RAket and RAvijn, but not for the set phonological syllable (bə)s , and the phoneticencoder can prepare the first phonetic syllablethat included raKET and loKET.
According to the WEAVER model, this se- [bə]. The remainder of the phonetic plan hasto be computed during the trial itself. Thus, forriality phenomenon reflects the suspension-re-
sumption mechanism that underlies the incre- polymorphemic words also initial morphemescan be prepared.mental planning of speech. Assume the re-
sponse set consists of raket, ravijn, and soOVERVIEW OF THE EXPERIMENTSforth (i.e., the first syllable is shared). Before
the beginning of a trial, the morphological en- Experiment 1 uses the implicit priming par-adigm to test the model’s prediction that thecoder can do nothing, the phonological en-
coder can construct the first phonological syl- constituent morphemes of compounds areplanning units in speech production. If effectslable (rɑ)s , and the phonetic encoder can re-
cover the first phonetic syllable [rɑ]. When of morphological complexity are obtained,this means that the paradigm is sensitive tothe prompt lucht is given, the morphological
encoder will retrieve »raket…. Segmental spell- morphological structure. It would corroborateevidence from speech errors that nominalout makes available the segments of this mor-
pheme, which includes the segments of the compounds are assembled out of their compo-nents during production.second syllable. The phonological and pho-
netic encoders can start working on the second The experiment tests whether the facilita-tory effect from a shared segmental string thatsyllable. In the heterogeneous condition (ra-
ket, bewijs, etc.), nothing can be prepared. constitutes a morpheme is larger than froma segmental string that does not constitute aThere will be no morphological encoding, no
phonological encoding, and no phonetic en- morpheme. For example, consider Dutch re-sponses that share the syllable bij. For mono-coding. In the end-homogeneous condition
(raket, loket, etc.), nothing can be done either. morphemic words such as bijbel (bible) con-sisting of the morpheme »bijbel…, only pho-Although the second syllable is known, the
phonological word cannot be computed be- nological preparation is possible. In the homo-geneous condition, (bεi)s and [bεi] will havecause the remaining segments are to the left
of the suspension point. In the model, this been planned for bijbel before the beginningof a trial, and »bijbel… and the second syllablemeans that the process has to go to the initial
segments of the word, which amounts to re- will be planned during the trial itself. In theheterogeneous condition, the whole word hasstarting the whole process. Thus, a facilitatory
effect will be obtained for the homogeneous to be planned during the trial. By contrast, forpolymorphemic words such as bijrol (support-condition relative to the heterogeneous condi-
tion for the begin condition only. Computer ing role) consisting of the morphemes »bij…and »rol…, additional morphological prepara-simulations of the experiments of Meyer
(1990) can be found in Roelofs (1994, submit- tion is possible. If »bij…, (bεi)s , and [bεi] havebeen planned for bijrol before the beginningted-a).
The suspension-resumption mechanism can of a trial in the homogeneous condition, »rol…can be selected during the trial itself, and thealso be applied to the production of polymor-
AID JML 2453 / a004$$$$83 11-07-96 07:33:33 jmlas AP: JML
861PRODUCTION OF POLYMORPHEMIC WORDS
second syllable can be computed. In the het- tion is due to the suspension-resumptionmechanism, then the number of times the verberogeneous condition, however, »bij… has to be
selected first, before »rol… and its segments can is produced should be irrelevant.In Experiments 1 through 4, there is notbe selected so that the second syllable can be
computed. Thus, in case of a polymorphemic only shared morphology but also semanticoverlap. Experiment 5 addresses the influenceword such as bijrol, additional morphological
preparation is possible before the beginning of semantic overlap per se in the implicit prim-ing paradigm by comparing the production ofof a trial. Consequently, extra facilitation
should be obtained. Thus, the facilitatory ef- monomorphemic words in semantically ho-mogeneous and heterogeneous sets.fect for bij in bijrol (consisting of the mor-
phemes »bij… and »rol…) should be larger thanthe effect for bij in bijbel (»bijbel…). EXPERIMENT 1
In Experiment 2, the effect of sharing aMethod
noninitial constituent morpheme is assessedfor nominal compounds. An end-homoge- Materials. The Dutch stimulus materials
consisted of two practice sets and twelve ex-neous set in the experiment consists, for exam-ple, of bijrol (»bij…»rol…), koprol (»kop…»rol…), perimental sets of three word pairs each. All
response words were disyllabic. There weredeegrol (»deeg…»rol…). Under the seriality as-sumption, it should not be possible to prepare six different homogeneous sets and six differ-
ent heterogeneous sets. Following Meyera noninitial morpheme (e.g., »rol… in bijrol).Thus, whereas the model predicts a facilitatory (1990, 1991), I will refer to the homogeneity
variable as Context. In the homogeneous con-effect from homogeneity for the initial mor-phemes of compounds, in this second experi- dition, the response words shared the first syl-
lable, whereas in the heterogeneous conditionment no facilitatory effect should be obtained.In Experiments 3 and 4, prefixed verbs have there was no such overlap. Hereafter, I will
refer to the critical part of a response as theto be produced. The base verbs of prefixedverbs are independent phonological words, Fragment. In half of the homogeneous sets
the shared syllable constituted a morpheme,but prefixes such as »ver…, »be…, and »ont… arenot. In Experiment 3, the effect of sharing the whereas in the other half of the homogeneous
sets it did not. Hereafter, I will refer to thisprefix is assessed, and Experiment 4 examinesthe effect of sharing the base. Under the serial- morphemic variable as Status (i.e., Pseudo
versus Real). The materials were obtained byity assumption, it should be possible to pre-pare the prefix but not the base. Thus, the an exhaustive search of the CELEX lexical
database (Baayen, Piepenbrock, & van Rijn,model predicts a facilitatory effect from ho-mogeneity for the prefixes but not for the 1993). Table 1 lists the materials of the experi-
ment.bases. By using trisyllabic verbs with mono-syllabic prefixes and disyllabic bases, this pre- Design. The experiment consisted of 12 ex-
perimental blocks administered consecutively.diction receives a strong test. Long word frag-ments will typically take longer to encode than Half of the blocks included a heterogeneous
set and half a homogeneous set. The first 6short fragments (Meyer, 1990, 1991). There-fore, the advantage from preparing long frag- blocks in the experiment were made up by
three homogeneous sets followed by three het-ments should be greater than from preparingshort fragments. However, contrary to this, the erogeneous ones (subject groups A and B) or
three heterogeneous sets followed by three ho-model predicts a facilitatory effect for theshort prefixes but no such effect for the long mogeneous ones (subject groups C and D).
The first syllable of the responses constitutedbases. Furthermore, with repeated productionof the prefixed verbs, the bases still should a morpheme (groups A and C) or it did not
(groups B and D). In a block, each of the threenot yield facilitation. If the absence of facilita-
AID JML 2453 / a004$$$$83 11-07-96 07:33:33 jmlas AP: JML
862 ARDI ROELOFS
TABLE 1
MATERIALS OF EXPERIMENT 1: PSEUDO VERSUS REAL MORPHEMES
Pairs
Pseudo
religie–bijbel graaf–hertog teen–nagel(religion–bible) (count–duke) (toe–nail)haast–bijna brein–hersens volk–natie(near–almost) (intellect–brain) (people–nation)spoor–bijster schaap–herder dicht–nader(trail–loss) (sheep–sheppard) (near–further)
Real
studie–bijvak boek–herdruk mond–nasmaak(study–subsidiary subject) (book–reprint) (mouth–after-taste)toneel–bijrol oorsprong–herkomst echo–nagalm(stage–supporting role) (source–origin) (echo–reverberation)long–bijnier steiger–herbouw herfst–najaar(lung–kidney) (scaffolding–rebuilding) (fall–autumn)
Homogenous Heterogeneous
Set 1 Set 2 Set 3 Set 1 Set 2 Set 3
Pseudo
BIJbel HERtog NAgel bijbel bijna bijsterBIJna HERder NAtie hersens herder hertogBIJster HERsens NAder nader nagel natie
Real
BIJvak HERdruk NAsmaak bijvak bijrol bijnierBIJrol HERkomst NAgalm herkomst herbouw herdrukBIJnier HERbouw NAjaar najaar nasmaak nagalm
Note. An approximate English translation of the words is given in parentheses.
pairs occurred randomly eight times. Thus, (NEC Multisync30) and a microphone (Senn-heisser ME40). After the subject had read thethere were 24 trials within a block. A pair was
not repeated on adjacent trials. In the next 6 instructions, two practice blocks (a homoge-neous and a heterogeneous one with the sameblocks, the remaining six homogeneous and
heterogeneous sets were presented. Again, the structure as an experimental block, but withdifferent items) were administered followedfirst syllable of the responses constituted a
morpheme (groups B and D) or it did not by the 12 experimental blocks. In the learningphase before each block, the three word pairs(groups A and C). The order of homogeneous
and heterogeneous sets was counterbalanced of a set were presented on the screen. As soonas the subject indicated having studied theacross subjects.
Procedure and apparatus. The subjects pairs sufficiently, the experimenter started thetest phase. The structure of a trial was as fol-were tested individually. They were seated in
a quiet room in front of a computer screen lows. First, the subject saw a warning signal
AID JML 2453 / a004$$$$83 11-07-96 07:33:33 jmlas AP: JML
863PRODUCTION OF POLYMORPHEMIC WORDS
TABLE 2(an asterisk) for 500 ms. Next, the screen wascleared for 500 ms, followed by the display MEAN PRODUCTION LATENCIES IN MILLISECONDS AND
ERROR PERCENTAGES (IN PARENTHESES) FOR EXPERI-of the prompt for 1500 ms. The asterisk andMENT 1prompt were presented in white on a black
background. Finally, before the start of theContext
next trial there was a blank interval of 500ms. Thus, the total duration of a trial was 3 Status Homogeneous Heterogeneous Diffs. The experiment was controlled by a Hermac
Pseudo 618 (2.7) 648 (0.4) 030386 SX computer.Real 625 (3.4) 699 (2.4) 074Analyses. After each trial, the experimenterTotal 621 (3.0) 674 (1.4) 053
coded the response for errors. Experimentalsessions were recorded on audio tape by aSony DTC55 DAT recorder. The recordingscontained the subjects’ speech and tones indi- dition than in the heterogeneous condition (the
by-subject standard errors of the means were,cating the onset of the prompt (1 kHz) andthe moment of the triggering of the voice key respectively, 7.9 and 7.7 ms). Most impor-
tantly, the facilitatory effect of homogeneity(2.5 kHz). These tones were also heard by theexperimenter (via closed headphones) at each was larger when the shared fragment consti-
tuted a morpheme (the Real condition) thantrial. The recordings were consulted after theexperiment when the experimenter was in when it did not (the Pseudo condition).
The statistical analyses yielded a main ef-doubt about whether a response was fully cor-rect. Four types of incorrect responses were fect of Context (F1(1,8) Å 41.29, MSe Å 7154,
p õ .001; F2(1,12) Å 189.92, MSe Å 130, pdistinguished. First, a subject might have pro-duced a wrong response word. Second, the õ .001), Status (F1(1,8) Å 10.28, MSeÅ 8967,
p õ .01; F2(1,12) Å 5.31, MSe Å 1446, p õresponse might have exhibited a disfluency,that is, the subject stuttered, paused within the .04), and Fragment (F1(2,16) Å 21.60, MSe Å
4588, põ .001; F2(2,12) Å 5.71, MSe Å 1446,utterance, or repaired the utterance. Third, thevoice key might have been triggered by a non- p õ .02). Most importantly, Context inter-
acted with Status (F1(1,8) Å 11.06, MSe Åspeech sound (noise in the environment or asmacking sound produced by the lips or 4798, p õ .01; F2(1,12) Å 34.10, MSe Å 130,
p õ .001): The effect of Context was largertongue). Fourth, the subject might have failedto respond within a time-out period of 1500 for the real morphemes than for the pseudo
morphemes. There was no triple interactionms. Incorrect responses were excluded fromthe statistical analysis of the production laten- between Context, Status, and Fragment
(F1(2,16) õ 1, MSe Å 4291, p ú .80; F2(2,12)cies.The production latencies and error rates õ 1, MSe Å 130, p ú .62).
Table 2 also gives the overall error ratewere submitted to by-subject and by-itemanalyses of variance with Context, Status, and (wrong responses and disfluencies) for the ho-
mogeneous and heterogeneous conditions.Fragment as repeated measures factors.Subjects. The experiment was conducted The percentages for the conditions with the
pseudo and real morphemes were 1.5 and 2.9.with 12 paid subjects from the pool of theMax Planck Institute. All subjects were native The total percentage of time-outs was 0.4 for
the homogeneous condition and 0.8 for thespeakers of Dutch.heterogeneous condition, and the percentages
Results and Discussion of false triggering of the voice-key were, re-spectively, 0.8 and 0.6.Table 2 gives the mean production latencies
as a function of Context and Status. The words Statistical analyses of the errors yielded amain effect of Context in the by-subject analy-were produced faster in the homogeneous con-
AID JML 2453 / a004$$$$83 11-07-96 07:33:33 jmlas AP: JML
864 ARDI ROELOFS
sis (F1(1,8) Å 6.31, MSe Å 0.23, p õ .04; would lead to such a general increase in laten-cies (Levelt, 1989). Word frequency is a moreF2(1,12) Å 4.64, MSe Å 0.12, p ú .05), but
no effect of Status (F1(1,8) Å 2.34, MSe Å likely cause (cf. Jescheniak & Levelt, 1994).The simple words were of higher frequency0.62, p ú .15; F2(1,12) Å 1.31, MSe Å 0.12,
p ú .27). Context did not interact with Status than the morphologically complex words(means are 67 and 3 per million in the CELEX(F1(1,8)Å 4.43, MSeÅ 0.33, pú .06; F2(1,12)
Å 4.64, MSe Å 0.12, p ú .05). database). Given that the overlap concernedthe first syllable only and given that each wordThe main effect of Context for the errors
might indicate that there is a speed–accuracy occurred in both the homogeneous and hetero-geneous condition, frequency cannot accounttradeoff in the data (e.g., Luce, 1986). On av-
erage, in the whole experiment (i.e., on the for the difference in facilitatory effect betweenthe pseudo and real conditions.288 trials for a subject) a single subject made
four errors in the homogeneous condition and In conclusion, a facilitatory effect is ob-tained from homogeneity of the first syllabletwo errors in the heterogeneous condition.
Thus, one might argue that subjects were of disyllabic words. The effect is larger whenthe syllable constitutes a morpheme than whenfaster in the homogeneous condition than in
the heterogeneous condition at the cost of it does not: a morpheme preparation effect.The results of the experiment show that themaking more errors. Under this tradeoff hy-
pothesis, the sizes of the Context effect for the experimental paradigm is sensitive to morpho-logical structure. The outcome supports thelatencies and the errors should be positively
correlated. The difference in number of errors idea that the component morphemes of com-pounds are planning units in speech produc-between the homogeneous and the heteroge-
neous condition should be larger when the dif- tion, confirming evidence from speech errors.Speech errors suggest that compounds andference between the corresponding latencies
(i.e., the facilitatory effect) is larger. This, prefixed words are assembled out of their mor-phemes in production. Furthermore, the resultshowever, was not the case. The Pearson prod-
uct–moment correlation between the error ef- of Experiment 1 show that the implicit primingparadigm is sensitive to morphological structure.fect and the latency effect was by-subjects r
Å 0.02 (p ú .94) and by-items r Å 0.46 (p The next three experiments test whether succes-sive morphemes of a compound or prefixedÅ .05). Thus, contrary to the tradeoff hypothe-
sis, the correlation between the effect sizes for word have to be produced in serial order. Exper-iment 1 showed that the initial morpheme ofthe latencies and errors was negative rather
than positive (if anything). In short, the results compounds can be prepared. Experiment 2 looksat the effect of sharing a noninitial morphemeof the statistical analyses of the errors do not
change the interpretation of the production la- for these compounds. Under the seriality as-sumption, it should not be possible to preparetencies.
The latency difference between the pseudo such noninitial morpheme. In the next two ex-periments, prefixed verbs have to be produced.and real conditions emerges for the heteroge-
neous condition but not for the homogeneous In Experiment 3, the effect of sharing the prefixis assessed for these prefixed verbs, and Experi-condition. This may seem odd if the homoge-
neous context is supposed to produce facilita- ment 4 examines the effect of sharing the base.Under the seriality assumption, it should be pos-tion. Direct comparisons of this sort are, how-
ever, problematic, because they involve com- sible to prepare the prefix but not the base. Thus,the model predicts a facilitatory effect from ho-parisons between different words. It might
simply be that the morphologically complex mogeneity for the prefixes but not for the bases.Furthermore, with repeated production of thewords have longer average latencies than do
the morphologically simple words. There is words, sharing noninitial morphemes should stillnot yield a facilitatory effect.little reason to expect that complexity itself
AID JML 2453 / a004$$$$84 11-07-96 07:33:33 jmlas AP: JML
865PRODUCTION OF POLYMORPHEMIC WORDS
METHOD of the Max Planck Institute. All subjects werenative speakers of Dutch.Materials
The Dutch materials in each experiment EXPERIMENT 2consisted of two practice sets and six experi- In this second experiment, subjects have tomental sets of three word pairs each. There produce nominal compounds (e.g., bijrol »bij…were three different homogeneous and hetero-
»rol…, koprol »kop…»rol…, deegrol »deeg…»rol…).geneous sets. In the homogeneous condition, The experiment assesses the effect of sharingthe responses shared morphemes, whereas in a noninitial constituent morpheme (in the ex-the heterogeneous condition they did not. ample, »rol…). According to the model, itThe materials were obtained by an exhaus- should not be possible to prepare such a noni-tive search of the CELEX lexical database nitial morpheme of a word. Thus, in contrast(Baayen et al., 1993). to Experiment 1, the model predicts no facili-
tation from homogeneity of the morphemes.DesignTable 3 lists the materials.Each experiment consisted of 18 experimen-
tal blocks administered consecutively. Half of Results and Discussionthe blocks included a heterogeneous set and half
The compounds were produced slightlya homogeneous set. The first 6 blocks in theslower in the homogeneous condition than inexperiment (first Repetition) were made up bythe heterogeneous condition. The mean pro-the three homogeneous sets followed by theduction latencies for the homogeneous andthree heterogeneous ones (one-half of the sub-heterogeneous conditions were respectivelyjects) or the three heterogeneous sets followed700 and 687 ms (the by-subject standard errorsby the three homogeneous ones (the other halfof these means were, respectively, 6.2 and 6.3of the subjects). Thus, after the first 3 blocks, ams). The statistical analysis showed that theresubject had seen all nine pairs. In a block, eachwas no significant difference between theseof the three pairs occurred randomly six times,conditions (F1(1,10) Å 1.77, MSe Å 14823, pso there were 18 trials within a block. A pairú .21; F2(1,6) Å 1.34, MSe Å 1630, p ú .29).was not repeated on adjacent trials. In the nextThe production latencies decreased with repe-6 blocks, the homogeneous and heterogeneoustition (F1(2,20) Å 9.80, MSe Å 14739, p õsets were presented for the second time, but now.001; F2(2,12)Å 70.62, MSe Å 170, põ .001).in a different order of blocks (second Repeti-The by-subject analysis yielded a main effecttion). Again, three heterogeneous blocks fol-of Fragment (F1(2,20) Å 12.41, MSe Å 15946,lowed the three homogeneous ones or vicep õ .001; F2(2,6) Å 3.24, MSe Å 5091, p úversa, depending on the group of subjects. The.11). Context did not interact with Repetitionsame held for the last 6 blocks (third Repetition).(F1(2,20) õ 1, MSe Å 7677, p ú .64; F2(2,12)The order of homogeneous and heterogeneousõ 1, MSe Å 430, p ú .52) or Fragmentsets was counterbalanced across subjects.(F1(2,20) Å 1.12, MSe Å 5731, p ú .34;
Procedure, Apparatus, and Analyses F2(2,6) õ 1, MSe Å 1630, p ú .73). Thus,with repeated production of the compounds,This was the same as in Experiment 1. Thestill no effect of Context is obtained.production latencies and error rates were sub-
The overall error rate for the homogeneousmitted to by-subject and by-item analyses ofand heterogeneous conditions was 2.5 andvariance with Context, Repetition, and Frag-1.0%, respectively. The total percentage ofment as repeated measures factors.time-outs was 0.7 for the homogeneous condi-
Subjects tion and 0.9 for the heterogeneous condition,and the percentages of false triggering of theEach experiment was conducted with a dif-
ferent group of 12 paid subjects from the pool voice-key were, respectively, 1.4 and 1.0. The
AID JML 2453 / a004$$$$84 11-07-96 07:33:33 jmlas AP: JML
866 ARDI ROELOFS
TABLE 3
MATERIALS OF EXPERIMENT 2: NOMINAL COMPOUNDS SHARING THE SECOND CONSTITUENT MORPHEME
Pairs
pannekoek–deegrol gevoel–indruk familie–stamboom(pancake–dough roll) (feeling–impression) (family–pedigree)gymnastiek–koprol foto–afdruk trein–spoorboom(gymnastics–somersault) (photo–print) (train–barrier)figurant–bijrol boek–herdruk gewicht–hefboom(super numerary–supporting role) (book–reprint) (weight–power lift)
Homogeneous Heterogeneous
Set 1 Set 2 Set 3 Set 1 Set 2 Set 3
deegROL inDRUK stamBOOM deegrol koprol bijrolkopROL afDRUK spoorBOOM afdruk herdruk indrukbijROL herDRUK hefBOOM hefboom stamboom spoorboom
Note. An approximate English translation of the words is given in parentheses.
statistical analyses of the errors did not yield (i.e., »halen…, »lopen…, »schieten…). The prefixesused are the most productive ones in Dutchsignificant effects.
The absence of a facilitatory effect from (Lieber & Baayen, 1993). In the homogeneouscondition, the response words shared the pre-homogeneity suggests that in producing nomi-
nal compounds, a noninitial morpheme cannot fix or the base, whereas in the heterogeneouscondition there was no such overlap. Allbe planned in advance. Thus, the results for
these polymorphemic words agree with the prompts were nouns and all responses wereprefixed verbs. The prompt named a typicalfindings obtained by Meyer (1990) for mono-
morphemic words. Furthermore, with re- theme/patient for the verb so that the associa-tion between prompt and response would bepeated production of the compounds, shared
noninitial morphemes still do not yield facili- natural and easy to remember. Table 4 liststhe materials of the experiments.tation. This suggests that preparation of noni-
nitial morphemes cannot be learned in the In Experiment 3, subjects have to produceprefixed verbs. The experiment assesses thecourse of an experiment. The results confirm
the prediction of the model. effect of sharing the prefix. According to themodel, it should be possible to prepare the
EXPERIMENTS 3 AND 4 initial morpheme of an utterance. Thus, themodel predicts a facilitatory effect from ho-In the next two experiments, prefixed verbsmogeneity of the prefixes.have to be produced. In Experiment 3, the
effect of sharing the prefix is assessed forExperiment 3these prefixed verbs, and Experiment 4 exam-
ines the effect of sharing the base. In each Results and discussion. The prefixed wordswere produced faster in the homogeneous con-experiment, there were three different homo-
geneous sets and three different heterogeneous dition than in the heterogeneous condition.The mean production latencies for the homo-sets. The nine experimental response words
consisted of all possible combinations of three geneous and heterogeneous conditions wererespectively 657 and 684 ms (the by-subjectprefixes (i.e., »be…, »ont…, »ver…) and three bases
AID JML 2453 / a004$$$$84 11-07-96 07:33:33 jmlas AP: JML
867PRODUCTION OF POLYMORPHEMIC WORDS
TABLE 4
MATERIALS OF EXPERIMENTS 3 AND 4: PREFIXED VERBS SHARING THE PREFIX (EXPERIMENT 3)OR THE BASE VERB (EXPERIMENT 4)
Pairs
afstand–belopen militair–beschieten resultaat–behalen(distance–walk) (soldier–shoot at) (result–obtain)paspoort–verlopen kleur–verschieten gebeuren–verhalen(passport–expire) (color–fade) (event–tell about)straf–ontlopen naam–ontschieten vriend–onthalen(punishment–escape) (name–slip memory) (friend–regal with)
Homogeneous Heterogeneous
Set 1 Set 2 Set 3 Set 1 Set 2 Set 3
Prefix
ONTlopen BElopen VERlopen ontlopen ontschieten onthalenONTschieten BEschieten VERschieten beschieten behalen belopenONThalen BEhalen VERhalen verhalen verlopen verschieten
Base
beLOPEN beSCHIETEN beHALEN belopen verlopen ontlopenverLOPEN verSCHIETEN verHALEN verschieten ontschieten beschietenontLOPEN ontSCHIETEN ontHALEN onthalen behalen verhalen
Note. An approximate English translation of the words is given in parentheses.
standard errors of these means were 6.6 and tion and 1.5 for the heterogeneous condition,and the percentages of false triggering of the6.6 ms). Statistical analysis showed that the
27 ms effect of Context was significant voice-key were, respectively, 0.4 and 0.7. Thestatistical analyses of the errors yielded a sig-(F1(1,10) Å 10.36, MSe Å 11126, p õ .009;
F2(1,6) Å 20.16, MSe Å 477, p õ .004). The nificant effect for Repetition only (i.e., thenumber of errors decreased with repetition).production latencies decreased with repetition
(F1(2,20) Å 4.76, MSe Å 13238, p õ .02; The facilitation effect from Context for theprefixes is smaller than the effect for the initialF2(2,12) Å 13.77, MSe Å 381, p õ .001). The
by-subject analysis yielded a main effect of compound morphemes in Experiment 1. Acomparison of this sort is, however, problem-Fragment (F1(2,20) Å 20.08, MSe Å 9097, p
õ .001; F2(2,6) Å 3.24, MSe Å 4697, p ú atic, because it involves a direct comparisonbetween different fragments. For example, the.11). Context did not interact with Repetition
(F1(2,20) õ 1, MSe Å 6088, p ú .47; F2(2,12) compound morphemes are independent pho-nological words, but the prefixes are not. Fur-Å 1.69, MSe Å 231, p ú .22) but did interact
with Fragment in the by-subject analysis thermore, the segments differ.In conclusion, the results suggest that in(F1(2,20) Å 4.21, MSe Å 4567, p õ .03;
F2(2,6) Å 3.36, MSe Å 477, p ú .10). producing prefixed verbs, the prefix can beplanned in advance. Thus, the results for theseThe overall error rate for the homogeneous
and heterogeneous conditions was 3.9 and 2.3 polymorphemic words agree with the findingsobtained by Meyer (1990) for monomorphe-percent, respectively. The total percentage of
time-outs was 0.7 for the homogeneous condi- mic words and agree with the findings from
AID JML 2453 / a004$$$$84 11-07-96 07:33:33 jmlas AP: JML
868 ARDI ROELOFS
Experiment 1. The results confirm the predic- ment 1 provided evidence for phonologicalpreparation of monomorphemic and polymor-tion by the model.
In Experiment 4, subjects have to produce phemic words. If morphological preparationof the bases is possible, then a facilitatorythe same prefixed verbs as in the previous
experiment. The experiment assesses the ef- effect should have been obtained. However,no facilitatory effect is obtained for the basesfect of sharing the base verb. According to the
model, it should not be possible to prepare of prefixed words. This may imply that thelexical entries of prefixed words have no inter-noninitial morphemes of an utterance. Thus,
there should be no facilitatory effect from ho- nal morphological structure. However, underthis assumption, the evidence from speech er-mogeneity of the bases. Furthermore, with re-
peated production of the prefixed verbs, the rors for morphological structure in the formentries of prefixed words is difficult to explainbases should still not yield facilitation.(e.g., Dell, 1986; Stemberger, 1985a, 1985b).
Experiment 4 In short, the results of the experiment suggestthat the noninitial bases of prefixed words can-Results and discussion. The mean production
latencies for the homogeneous and heteroge- not be planned in advance of the prefixes.Thus, the results for these polymorphemicneous conditions were respectively 667 and 661
ms (the by-subject standard errors were 5.5 and words agree with the findings obtained byMeyer (1990) for monomorphemic words.5.5 ms). The statistical analysis showed that
there was no significant difference between Furthermore, with repeated production of theprefixed verbs, bases still do not yield facilita-these conditions (F1(1,10) õ 1, MSe Å 4825, p
ú .35; F2(1,6) õ 1, MSe Å 721, p ú .49). The tion. This suggests that preparation of nonini-tial morphemes cannot be learned in theproduction latencies decreased with repetition
(F1(2,20) Å 30.28, MSe Å 7600, p õ .001; course of an experiment. The results confirmthe prediction by the model.F2(2,12) Å 72.94, MSe Å 263, põ .001). There
was no main effect of Fragment (F1(2,20) Å Given that there was a significant (by-sub-ject) Context by Fragment interaction in Ex-2.20, p ú .13; F2(2,6) õ 1, MSe Å 6745, p ú
.76). Context did not interact with Repetition periment 3, it would be worthwhile to deter-mine whether the effect of the variable Linear(F1(2,20) õ 1, MSe Å 11328, p ú .80; F2(2,12)
Å 1.10, MSe Å 186, p ú .36) or Fragment order of overlapping morphemes (begin ver-sus end) is reliable in a between-experiment(F1(2,20) õ 1, MSe Å 7277, p ú .90; F2(2,6)
õ 1, MSe Å 721, p ú .95). Thus, with repeated statistical analysis, as is implied by the inter-pretation of the results. The interaction be-production of the prefixed words, still no effect
of Context is obtained. tween Context and Linear order of overlap-ping morphemes was significant (F1(1,20) ÅIn the experiment, the overall error rate for
the homogeneous and heterogeneous condi- 10.37, MSe Å 7975, p õ .004; F2(1,12) Å11.51, MSe Å 599, p õ .005). Furthermore,tions was 2.0 and 1.5%, respectively. The total
percentage of time-outs was 0.7 for the homo- this interaction did not depend on Fragment(F1(2,40) Å 1.23, MSe Å 5922, p ú .30;geneous condition and also 0.7 for the hetero-
geneous condition, and the percentages of F2(1,12) Å 1.02, MSe Å 599, p ú .39). So,the between-experiment analyses confirm thefalse triggering of the voice-key were, respec-
tively, 0.6 and 0.4. The statistical analyses of earlier interpretation of the results.the errors did not yield significant effects.
EXPERIMENT 5The results of Experiment 4 suggest that inproducing prefixed verbs, the base cannot be This final experiment addresses the influ-
ence of semantic overlap in the implicit prim-planned in advance. If phonological prepara-tion of the base is possible, then facilitation ing paradigm. The argument in Experiment 1
was that morphological overlap per se pro-should have been obtained. Recall that Experi-
AID JML 2453 / a004$$$$84 11-07-96 07:33:33 jmlas AP: JML
869PRODUCTION OF POLYMORPHEMIC WORDS
duced the extra facilitation. The difference be- trein; hoef–paard, mouw–jas, spaak–fiets.Thus, as in the previous experiments, thetween initial syllables constituting a mor-
pheme (e.g., bij in bijrol) and syllables that prompts came from different semantic do-mains.do not (e.g., bij in bijbel) was interpreted to
support the idea that morphemes are planning Design, procedure, apparatus, and analy-ses. This was the same as in Experiments 2units in speech production. However, the se-
mantic overlap among the response words to 4, except that the responses in the homoge-neous sets now denote semantic categoryseemed to be greater when a morpheme is
shared than when it is not. Therefore, one may members.Subjects. The experiment was run with sixargue that the results reflect semantic similar-
ity and do not mean that polymorphemic new subjects from the pool of the Max PlanckInstitute.forms are morphologically analyzed in the
mental lexicon. Of course, this view does notResults and Discussionexplain why the presumed facilitatory effect
from semantic overlap depends on serial posi- The means for the semantically homoge-neous and heterogeneous conditions were, re-tion. In Experiments 2 and 4, words sharing
noninitial morphemes also shared part of their spectively, 641 and 623 ms (the by-subjectstandard errors were 7.4 and 7.4 ms). Thesemantics but did not yield a facilitatory ef-
fect. The explanation of why the facilitatory small inhibitory effect of 18 ms from semanticsimilarity was statistically significant (F1(1,4)effect of semantic similarity depends on linear
position would require the notion of mor- Å 10.48, MSe Å 808, p õ .03; F2(1,6) Å 9.51,MSeÅ 445, põ .02). The production latenciespheme and the idea of successive encoding of
morphemes. Still, it would be important to decreased with repetition (F1(2,8) Å 20.86,MSe Å 658, p õ .001; F2(2,12) Å 21.68, MSedirectly address the role of semantic overlap.
In the fifth experiment, subjects have to pro- Å 317, p õ .001). The subject-analysisyielded a main effect of Semantic domainduce monomorphemic nouns. In the homoge-
neous sets, the nouns denote semantic cate- (F1(2,8) Å 4.97, MSe Å 1882, põ .04; F2(2,6)Å 1.15, MSe Å 4065, p ú .37). There wasgory members (e.g., dog, horse, mouse) and
in heterogeneous sets they do not. If semantic no interaction between Context and Repetition(F1(2,8) õ 1, MSe Å 632, p ú .59; F2(2,12)similarity is the cause of the difference be-
tween the real condition (e.g., bijrol) and the Å 1.05, MSe Å 170, p ú .38) nor betweenContext and Semantic domain (F1(2,8) Åpseudo condition (e.g., bijbel) in Experiment
1, then a facilitatory effect should be obtained 1.28, MSe Å 972, p ú .33; F2(2,6) Å 1.40,MSe Å 445, p ú .31).from semantic similarity in the current experi-
ment. The overall error rate for the homogeneousand heterogeneous conditions was 1.0 and 0.0
Method percent, respectively. The total percentage oftime-outs was 1.0 for the homogeneous condi-Materials. Three semantic domains were
used, namely animals, clothes, and means of tion and 0.0 for the heterogeneous condition,and the percentage of false triggering of thetransportation. The homogeneous sets were
kaas–muis (cheese–mouse), mand–hond voice-key was 0.0. There were no significanteffects.(basket–dog), and hoef–paard (hoof–horse);
mouw–jas (sleeve–coat), zool–schoen (sole– In conclusion, large semantic overlapamong the response words yields inhibitionshoe), and haar–pet (hair–cap); biels–trein
(sleeper– train), spaak–fiets (spoke–bike), instead of facilitation (although for somewords in Experiments 1–4 there was semanticand mast–boot (mast–boat). The heteroge-
neous sets were: kaas–muis, zool–schoen, overlap, no inhibition was obtained there pre-sumably because the overlap was not largemast–boot; mand–hond, haar–pet, biels–
AID JML 2453 / a004$$$$85 11-07-96 07:33:33 jmlas AP: JML
870 ARDI ROELOFS
enough). Thus, the finding in Experiment 1that initial syllables constituting a morpheme(e.g., bij in bijrol) produce more facilitationthan syllables that do not (e.g., bij in bijbel)cannot be due to semantic similarity. The ar-gument that morphological overlap per se isa significant cause of facilitation and that mor-phemes are therefore units of production plan-ning remains supported.
COMPUTER SIMULATIONS
The WEAVER model of form encoding inspeech production (Roelofs, 1992b, 1994,submitted-a) was taken as the theoreticalframework for the present paper because itis more explicit than other models about theparticular issues addressed. For example, Dell
FIG. 1. Memory representation of the word form of the(1986) adopted the hypothesis of morphologi- Dutch word bijrol in the model.cal decomposition, but explored his modelthrough computer simulation for monomor-phemic words (monosyllables and disyllables)
speech production have ignored certain as-only. The revised version of this model pro-pects of morphological complexity. One canposed in Dell (1988) has not been exploredonly speculate about how these models wouldthrough computer simulation, and has beenhandle the findings from the experiments inapplied to monomorphemic monosyllablesthe present paper.only. The same holds for the computer imple-
mentation of Schade and Berg’s (1992) ver-Overview of the WEAVER Model of Speechsion of Dell’s (1988) model. Similarly, Harley
Production(1993) implemented an interactive activationmodel of lexical access, but explored his Figure 1 illustrates the memory representa-
tion of the form of the Dutch word bijrol inmodel for monomorphemic monosyllablesonly. Finally, Stemberger (1985b) adopted the the WEAVER model. Recall that the form
lexicon is conceived of as a network. Theredecomposition assumption in his interactive-activation model, but did not explore his is a metrical part and a nonmetrical part. The
nonmetrical part consists of three layers ofmodel through computer simulation. More-over, many aspects of his model relevant for nodes: morpheme nodes, segment nodes, and
syllable program nodes. Morpheme nodesthe present paper are unclear. For example,although Stemberger explicitly assumed par- stand for roots and affixes. Morpheme nodes
are connected to the lemma. For example, theallel activation of words in a syntactic struc-ture (‘‘. . . all of the words in a clause and nodes »bij… and »rol… are connected to the
lemma of bijrol. A morpheme node points toall the segments of a word are selected at thesame time,’’ Stemberger, 1985a, p. 274), it is two major types of information, namely to its
canonical metrical structure and to the seg-not fully clear whether this assumption shouldbe extended to polymorphemic words. If par- ments that make up its underlying form. The
metrical structure describes an abstract group-allel encoding is extended to the morphologi-cal level, this would contradict the findings ing of syllables (s) into feet (S) and feet into
phonological words (v). (The foot level is om-from the current experiments. In short, ex-isting models of word-form encoding in mited from the figure for reasons of simplic-
AID JML 2453 / a004$$$$85 11-07-96 07:33:33 jmlas AP: JML
871PRODUCTION OF POLYMORPHEMIC WORDS
ity.) Importantly, it is not specified which seg- nucleus, coda) are assigned to the segmentsfollowing the syllabification rules of the lan-ments make up the syllables. The links be-
tween morpheme and segment nodes indicate guage. Thus, in the encoding of »bij…»rol…, the/r/ is made syllable onset, the /ɔ/ nucleus, andthe serial position of the segments within the
morpheme. Possible syllable positions (onset, the /l/ coda. The phonetic encoder selects thesyllable program nodes whose labeled linksnucleus, coda) of the segments are specified
by the links between segment nodes and sylla- to the segments correspond with the syllablepositions assigned to the segments. For exam-ble program nodes. For example, the network
specifies that /l/ is the coda of [rɔl] and the ple, [rɔl] is selected for the second phonologi-cal syllable of bijrol, because the link betweenonset of [lə]. These links are used in retrieving
an articulatory program for a syllable after the [rɔl] and /r/ is labeled onset, between [rɔl]and /ɔ/ nucleus, and between [rɔl] and /l/ coda.actual syllable positions of the segments have
been determined by the syllabification pro- In producing the plural form bijrollen (»bij…»rol…»en…, syllabified as (bεi)s(rɔ)s(lə)s), thecess.
Information is retrieved from the network /l/ will be syllabified with the plural mor-pheme »en…, and the syllable program [lə] willthrough spreading of activation. Encoding
starts when a morpheme node receives activa- be selected. The plural suffix in Dutch is notan independent domain of syllabificationtion from a lemma. Activation then spreads
through the network in a strictly forward fash- (compare the syllabification of bijrollen to thatof behalen, illustrated in the introductory sec-ion. Each node sends a proportion of its acti-
vation to its direct neighbors. There is also tion of this paper). Thus, the model providesfor syllabification across morpheme bound-spontaneous decay of activation. Since several
morphemes and their segments may be avail- aries. If the selection conditions of a syllableprogram node are met, the actual selection ofable at a particular moment in time, the encod-
ing algorithm has to select the relevant nodes the node at any moment in time is a randomevent. The probability of selecting a node atamong all the activated ones in order to syllab-
ify them. To accomplish this task, the form a particular moment in time is equal to theratio of its level of activation and the sum ofencoders follow simple selection rules. The
rules are implemented in a parallel distributed the activation levels of all syllable programsnodes in the network.manner. Attached to each node in the network,
there is a procedure that verifies the label on The simulations involved word-form en-coding up to the access of the phonetic sylla-the link between the node and a target node
one level up. A verification procedure is trig- bary. The mathematical equations for thespreading of activation, the selection ratio, andgered when the node’s activation level ex-
ceeds a threshold. The procedures may run in the expectation of the word-form encoding la-tency are as follows (cf. Roelofs, 1992a,parallel.
The morphological encoder selects the mor- 1992b, 1993, submitted-a). Activation spreadsaccording topheme nodes that are linked to a selected
lemma. Thus, »bij… and »rol… are selected forthe lemma of bijrol. The phonological encoder a(k, t / Dt) Å a(k, t)(1 0 d) / ∑
n
r a(n, t),selects the segments and the metrical structurethat are linked to the selected morphemenodes. The segments are associated to the syl- where a(k, t) is the activation level of node k
at point in time t, d is a decay rate (0 õ d õlable nodes within the metrical frame. Theassociation proceeds from the segment whose 1), and Dt is the duration of a time step (in
ms). The rightmost term denotes the amountlink is labeled first to the one labeled second,and so forth. In associating the segments to of activation k receives between t and t / Dt,
where a(n, t) is the output of neighbor n (equalthe metrical frame, syllable positions (onset,
AID JML 2453 / a004$$$$85 11-07-96 07:33:33 jmlas AP: JML
872 ARDI ROELOFS
to its level of activation). The factor r indi-cates the spreading rate.
The probability that a target node m will beselected at t õ T £ t / Dt given that it hasnot been selected at T £ t, and provided thatthe selection conditions for a node are met, isgiven by the ratio
p(selection m at t õ T
£ t / Dt É ¬ selection m at T £ t)
Å a(m, t)(i a(i, t)
.
The index i ranges over the syllable programnodes in the word-form network of a speaker.The selection ratio equals the hazard rate hm(s)of the process of the encoding of syllable m(up to the access of the syllabary) at time steps (cf. Luce, 1986, McGill, 1963; Townsend &Ashby, 1983), where t Å (s 0 1)Dt, and s Å1, 2, . . . The expected latency of word-formencoding up to the access of the syllabary,
FIG. 2. Results of the computer simulations.E(T), for disyllables is
Edi(T) Å ∑`
sÅ1
{f1(s) ∑s01
jÅ0
f2(s) / f2(s) ∑s01
jÅ0
f1(s) begin-homogeneous response sets consistedof bijrol (»bij…»rol…), bijnier (»bij…»nier…), and
/ f1(s)f2(s)}sDt, so forth (Status: Real), versus bijbel (»bijbel…),bijster (»bijster…), and so forth (Status:Pseudo). An end-homogeneous set consistedwhere f1(s) and f2(s) are the probability mass
functions of the encoding of the first syllable of bijrol (»bij…»rol…), koprol (»kop…»rol…), andso forth. The heterogeneous sets were createdand second syllable, respectively.
The parameter values were kept the same as by recombining the responses of different ho-mogeneous sets. The critical items were em-in the simulations reported in Roelofs (1994,
submitted-a). The spreading rate r within the bedded in a network which coded the formsof 50 other words randomly selected from theword-form stratum was 0.0120 [ms01], the de-
cay rate d was 0.0240 [ms01], and the size of CELEX lexical database (no embedding pro-duced the same simulation outcomes). Ad-the external input to the network extin was
0.1965 [ms01]. The activation threshold for vance knowledge about the form of the re-sponse words was simulated by completingthe triggering of a selection test was 1.5. The
duration of basic events such as the time for the morphological, phonological, and pho-netic encoding of the word form as far as pos-the activation to cross a link, the latency of a
selection test, and the syllabification time per sible before the beginning of a trial.Figure 2 shows the results of the simula-syllable equalledDt Å 25 ms. The completion
time of a morphemic procedure was twice as tions. Panel (a) shows the results of the simu-lation of Experiment 1. Sharing bij in bijrollong.
In the simulations of the experiments, the and bijbel yields a facilitatory effect. The fa-
AID JML 2453 / a004$$$$85 11-07-96 07:33:33 jmlas AP: JML
873PRODUCTION OF POLYMORPHEMIC WORDS
cilitatory effect for bij in bijrol (consisting of speech plan for the word onset is completed,and it would be maximal if articulation is initi-the morphemes »bij… and »rol…, the /MORPH
condition) is larger than the facilitatory effect ated upon completion of the plan for the wholeword. Various intermediate positions are pos-for bij in bijbel (»bijbel…, the0MORPH condi-
tion). This corresponds to the experimental sible.The absence of a facilitatory effect for thefindings. Panel (b) shows the results of the
simulation of Experiments 1 and 2 (similar to noninitial morphemes in Experiment 2 and 4may have a rather trivial explanation, namely,those of Experiments 3 and 4). A facilitatory
effect is obtained when the responses share that the experimental paradigm is not able topick up an effect of their preparation. If articu-the first morpheme (bijrol, bijnier, etc., the
BEGIN condition), but not when they share lation begins upon completion of the firstcompound syllable or the prefix syllable, anthe second morpheme (bijrol, koprol, etc., the
END condition). This corresponds to the ex- effect of preparation of the second compoundsyllable or the base syllables of the prefixedperimental findings. To conclude, the simula-
tions demonstrate that the WEAVER model words will not be picked up.2 There exist atleast two empirical arguments against theaccounts for the empirical phenomena.view that the second syllable is not involved
GENERAL DISCUSSION in the initiation of articulation.First, Meyer (1990, 1991) observed thatThe present work deals with the largely ne-
glected issue of the role of morphology in longer fragments yield larger facilitatory ef-fects than shorter fragments. In particular, thelexical access in speech production. The aim
of the current research was to examine facilitatory effect was larger when the re-sponses shared both the first and (part of) thewhether the successive morphemes of a word
are planned in serial order. In particular, I in- second syllable than when only the first sylla-ble was shared. This suggests that the secondvestigated whether the speech production sys-
tem can plan noninitial morphemes of a word syllable is involved in the initiation of articu-lation of a word.before initial ones. Five experiments were re-
ported that tested predictions of the WEAVER Second, I had subjects produce simplephrasal forms, in particular, Dutch verb-parti-model of form encoding that assumes seriality
in planning the production of polymorphemic cle combinations, which consist of two em-bedded phonological words corresponding towords. The results of the experiments sup-
ported the assumption of serial planning of the particle and the base verb (Roelofs, sub-mitted-b). In producing particle verbs in a par-morphemes. It was shown by computer simu-
lation that the model accounts for the findings. ticle-first infinitival form, the facilitatory ef-fect was larger when the responses shared bothIn this final section, I will discuss a number
of empirical and theoretical issues that are the particle syllable and the first base syllablethan when only the particle syllable wasraised by the research.shared. For example, the effect was larger for
Advance Planning and the Initiation of uitleven, uitlezen, uitlenen (uit is the particleArticulation syllable and le is the first base syllable) thanThe issue of advance planning pertains to
the relationship between the generation of the 2 Alternatively, one may argue that onset latencies aresimply insensitive to the encoding of elements that arespeech plan and articulation (e.g., Gordon &discontinuous with initial segments. Meyer and SchriefersMeyer, 1987; Sternberg, Knoll, Monsell, &(1991), however, obtained facilitatory effects from spokenWright, 1988). How much of the speech plansecond-syllable primes (presented via headphones) in pic-
for a word has to be completed before articula- ture naming. These results have been replicated with spo-tion is initiated? The critical fragment would ken primes in the implicit-priming paradigm (Roelofs,
submitted-a).be minimal if articulation is initiated when the
AID JML 2453 / a004$$$$85 11-07-96 07:33:33 jmlas AP: JML
874 ARDI ROELOFS
for uitzoeken, uitdraaien, uitgeven (sharing planning units in speech production, andtherefore support the decomposition viewthe particle uit only). This suggests that plan-
ning the second phonological word of these rather than the full listing view for thesewords. The effect of morphological structureverbs determines the initiation of articulation
rather than planning the first phonological is difficult to explain under the full listing hy-pothesis. According to this hypothesis, wordsword only.
In sum, if articulation is initiated upon com- such as bijrol and bijbel would have the samestructure at the morphemic level, namely »bij-pletion of (part of) the first syllable, overlap
that crosses the first syllable boundary should rol… and »bijbel… instead of »bij…»rol… and »bij-bel…. In conclusion, the experiment providesnot increase the facilitatory effect. The empiri-
cal evidence shows, however, that the facilita- evidence for a morphologically decomposedrepresentation of compounds, corroboratingtory effect becomes larger when overlap
crosses the first syllable boundary or the first existing evidence from speech errors.phonological word boundary. If (part of) the
What Is Serial in the Planning Process?first syllable would suffice, then the increaseof the facilitatory effect due to homogeneity The outcomes of Experiment 1 support the
decomposition hypothesis. In a decompositionof the second syllable remains unexplained.To conclude, a facilitatory effect from prepar- view, the morphemes of a word may be
planned in parallel or in a serial fashion. Theing the second compound morpheme and thebases of prefixed verbs could have been outcomes of Experiments 2, 3, and 4 sug-
gested that the successive morphemes of apicked up in Experiments 2 and 4.polymorphemic word are planned in serial
Decomposition versus Full Listing order.What exactly is serial in the planning pro-Under the full listing assumption, whole-
word forms rather than their constituent mor- cess? As the computer simulations demon-strated, the current findings do not imply thatphemes are represented in the mental lexicon
(e.g., »bijrol… and »bijbel…); under the decom- every aspect of planning the production ofpolymorphemic words is serial. For example,position assumption, the constituent mor-
phemes are stored (e.g., »bij…, »rol…, and »bij- seriality may hold for the application of pro-duction rules but not for the activation ofbel…). If constituent morphemes are not repre-
sented, they cannot be planning units in memory elements. The production rules in theWEAVER model give rise to rightward incre-speech production.
The WEAVER model assumes that syllabi- mentality, but the morphemes and the phone-mic segments of a word in memory are acti-fications of words are computed rather than
stored. Other models such as Dell’s (Dell, vated in parallel. For example, the representa-tions /b/, /εi/, /r/, /ɔ/, and /l/ for bijrol are1986, 1988) assume that syllabifications are
stored with words in memory. According to simultaneously activated. According to themodel, seriality reflects the fact that »bij… ismodels with an active syllabification process,
the syllabification process in word-form en- selected before »rol… and that the /b/ is selectedand syllabified before the /r/ (i.e., the workingcoding has to know the morphemic source of
the segments that are input to the process. The of production rules) rather than the activationof these elements in memory.process cannot blindly accept a string of seg-
ments and syllabify the segments without tak-CONCLUSIONing morpheme boundaries into account. This
implies that the lexical entries have to indicate The research reported in this paper providesevidence that the speech production systemmorpheme boundaries.
The outcomes of Experiment 1 suggest that has to plan the forms of the successive mor-phemes of a polymorphemic word in serialcomponent morphemes of compounds are
AID JML 2453 / a004$$$$86 11-07-96 07:33:33 jmlas AP: JML
875PRODUCTION OF POLYMORPHEMIC WORDS
frequency effects in speech production: Retrieval oforder. The results suggest that the system can-syntactic information and phonological form. Jour-not plan noninitial morphemes of a word be-nal of Experimental Psychology: Learning, Memory,
fore initial ones, even when these noninitial and Cognition, 20, 824–843.morphemes constitute fully fledged phonolog- JORDAN, M. I., & ROSENBAUM, D. A. (1989). Action. Inical words of their own. Thereby, the experi- M. I. Posner (Ed.), Foundations of cognitive science.
Cambridge, MA: MIT Press.ments confirm predictions of the WEAVERKEMPEN, G., & HUIJBERS, P. (1983). The lexicalizationmodel of word-form encoding in speech pro-
process in sentence production and naming: Indirectduction (Roelofs, 1992b, 1994, submitted-a).election of words. Cognition, 14, 185–209.
KEMPEN, G., & HOENKAMP, E. (1987). An incrementalREFERENCES procedural grammar for sentence formulation. Cog-
nitive Science, 11, 201–258.ANDERSON, J. R. (1983). The architecture of cognition.
LASHLEY, K. S. (1951). The problem of serial order inCambridge, MA: Harvard Univ. Press.
behavior. In L. A. Jeffress (Ed.), Cerebral mecha-BAAYEN, R. H., PIEPENBROCK, R., & VAN RIJN, H. (1993).
nisms in behavior. New York: Wiley.The CELEX Lexical database. (CD-ROM). Linguis-
LEVELT, W. J. M. (1989). Speaking: From intention totic Data Consortium, University of Pennsylvania,articulation. Cambridge, MA: MIT Press.Philadelphia, PA.
LEVELT, W. J. M. (1992). Accessing words in speechBOOIJ, G. E. (1983). Principles and parameters in prosodicproduction: Stages, processes and representations.phonology. Linguistics, 21, 249–280.Cognition, 42, 1–22.BOOIJ, G. E. (1995). The phonology of Dutch. Oxford:
LEVELT, W. J. M., & WHEELDON, L. (1994). Do speakersOxford Univ. Press.have access to a mental syllabary? Cognition, 50,BUTTERWORTH, B. (1983). Lexical representation. In B.239–269.Butterworth (Ed.), Language production: Vol. 2. De-
LIBERMAN, M., & PRINCE, A. (1977). On stress and lin-velopment, writing and other language processes.guistic rhythm. Linguistic Inquiry, 8, 249–336.London: Academic Press.
LIEBER, R., & BAAYEN, H. (1993). Verbal prefixes inDELL, G. S. (1986). A spreading-activation theory of re-Dutch: A study in lexical conceptual structure. Year-trieval in sentence production. Psychological Review,book of morphology, 51–78.93, 283–321.
LUCE, R. D. (1986). Response times: Their role in infer-DELL, G. S. (1988). The retrieval of phonological forms inring elementary mental organization. New York: Ox-production: Tests of predictions from a connectionistford University Press.model. Journal of Memory and Language, 27, 124–
MACKAY, D. G. (1978). Derivational rules and the inter-142.nal lexicon. Journal of Verbal Learning and VerbalGARRETT, M. F. (1975). The analysis of sentence produc-Behavior, 17, 61–71.tion. In G. H. Bower (Ed.), The psychology of learn-
MARSLEN-WILSON, W., TYLER, L. K., WAKSLER, R., &ing and motivation. New York: Academic Press.OLDER, L. (1994). Morphology and meaning in theGARRETT, M. F. (1980). Levels of processing in sentenceEnglish lexicon. Psychological Review, 101, 3–33.production. In B. Butterworth (Ed.), Language pro-
MCCARTHY, J., & PRINCE, A. (1990). Foot and word induction: Vol. 1. Speech and talk. London: Academicprosodic morphology: The Arabic broken plurals.Press.Natural Language and Linguistic Theory, 8, 209–GARRETT, M. F. (1988). Processes in language production.282.In F. J. Newmeyer (Ed.), Linguistics: The Cambridge
MCCARTHY, J., & PRINCE, A. (1993). Generalized align-survey (Vol. 3). Cambridge, MA: Harvard Universityment. Yearbook of Morphology, 79–154.Press.
MCGILL, W. J. (1963). Stochastic latency mechanisms.GOLDSMITH, J. (1990). Autosegmental and metrical pho-In R. D. Luce, R. R. Bush, & E. Galanter (Eds.),nology. Cambridge, MA: Blackwell.Handbook of mathematical psychology (Vol. 1). NewGORDON, P. C., & MEYER, D. E. (1987). Control of serialYork: Wiley.order in rapidly spoken syllable sequences. Journal
MEYER, A. S. (1990). The time course of phonologicalof Memory and Language, 26, 300–321.encoding in language production: The encoding ofHARLEY, T. A. (1993). Phonological activation of seman-successive syllables of a word. Journal of Memorytic competitors during lexical access in speech pro-and Language, 29, 524–545.duction. Language and Cognitive Processes, 8, 291–
MEYER, A. S. (1991). The time course of phonological309.encoding in language production: The phonologicalHENDERSON, L. (1985). Towards a psychology of mor-encoding inside a syllable. Journal of Memory andphemes. In A. W. Ellis (Ed.), Progress in the psy-Language, 30, 69–89.chology of language, Vol. 1. London: LEA.
JESCHENIAK, J-D., & LEVELT, W. J. M. (1994). Word MEYER, A. S. (1992). Investigation of phonological en-
AID JML 2453 / a004$$$$86 11-07-96 07:33:33 jmlas AP: JML
876 ARDI ROELOFS
coding through speech error analyses: Achievements, word-form encoding in speech production. Submit-ted for publication.limitations, and alternatives. Cognition, 42, 181–
211. ROELOFS, A. (submitted-b). Rightward incrementality inencoding simple phrasal forms in speech production:MEYER, A. S., & SCHRIEFERS, H. (1991). Phonological
facilitation in picture-word interference experiments: Verb-particle combinations. Submitted for publica-tion.Effects of stimulus onset asynchrony and types of
interfering stimuli. Journal of Experimental Psychol- SCHADE, U., & BERG, T. (1992). The role of inhibitionin a spreading-activation model of language produc-ogy: Learning, Memory, and Cognition, 17, 1146–
1160. tion. II. The simulational perspective. Journal of Psy-cholinguistic Research, 21, 435–462.NOOTEBOOM, S. G. (1969). The tongue slips into patterns.
In A. G. Sciarone, A. J. van Essen, & A. A. van SCHRIEFERS, H., ZWITSERLOOD, P., & ROELOFS, A. (1991).The identification of morphologically complex spo-Raad (Eds.), Nomen: Leyden studies in linguistics
and phonetics. The Hague: Mouton. ken words: Continuous processing or decomposi-tion? Journal of Memory and Language, 30, 26–47.PINKER, S., & PRINCE, A. (1988). On language and con-
nectionism: An analysis of a parallel distributed pro- SPENCER, A. (1991). Morphological theory. Cambridge,MA: Blackwell.cessing model of language acquisition. Cognition,
28, 73–193. STEMBERGER, J. P. (1985a). The lexicon in a model oflanguage production. New York: Garland Publish-POULISSE, N. (1989). The use of compensatory strategies
by Dutch learners of English. Doctoral dissertation, ing.STEMBERGER, J. P. (1985b). An interactive activationUniversity of Nijmegen.
ROELOFS, A. (1992a). A spreading-activation theory of model of language production. In A. W. Ellis (Ed.),Progress in the psychology of language (Vol. 1).lemma retrieval in speaking. Cognition, 42, 107–
142. London: LEA.STEMBERGER, J. P., & MACWHINNEY, B. (1986). Fre-ROELOFS, A. (1992b). Lemma retrieval in speaking: A
theory, computer simulations, and empirical data. quency and the lexical storage of regularly inflectedforms. Memory & Cognition, 14, 17–26.Doctoral dissertation, NICI Technical Report 92-08,
University of Nijmegen. STERNBERG, S., KNOLL, R. L., MONSELL, S., & WRIGHT,C. E. (1988). Motor programs and hierarchical orga-ROELOFS, A. (1993). Testing a non-decompositional the-
ory of lemma retrieval in speaking: Retrieval of nization in the control of rapid speech. Phonetica,45, 175–197.verbs. Cognition, 47, 59–87.
ROELOFS, A. (1994). On-line versus off-line priming of TOWNSEND, J. T., & ASHBY, F. G. (1983). Stochastic mod-eling of elementary psychological processes. Cam-word-form encoding in spoken word production. In
A. Ram & K. Eiselt (Eds.), Proceedings of the Six- bridge: Cambridge University Press.teenth Annual Conference of the Cognitive ScienceSociety, 772–777. Hillsdale, NJ: LEA. (Received April 13, 1995)
(Revision received September 18, 1995)ROELOFS, A. (submitted-a). The WEAVER model of
AID JML 2453 / a004$$$$86 11-07-96 07:33:33 jmlas AP: JML